The INESC-ID IWSLT07 SMT system

The INESC-ID IWSLT07 SMT System Jo˜ao V. Grac¸a , Diamantino Caseiro, Luisa Coheur Spoken Language Lab (L2F) Inesc-ID, Lisboa javg,dcaseiro,luisa.coheur@l2f.inesc-id.pt Abstract We present the machine translation system used by L2F from INESC-ID in the evaluation campaign of the International Workshop on Spoken Language Translation (2007), in the task of translating spontaneous conversations in the travel do- main from Italian to English. 1. Introduction This paper describes the machine translation system used by INESC-ID in its first participation on the evaluation cam- paign of the International Workshop on Spoken Language Translation 2007. We submitted translation results for manual and first-best transcriptions in the Italian-to-English language pair. The statistical machine translation system consists of a first-pass phrase-based system using moses machine transla- tion toolkit [1], followed by a reranking step. In section 2 we describe the system as well as the cor- pora and the baseline results; in section 3 we present several experiments we did in order to improve the results. Then, in section 4 we show the results we obtained. Finally, section 5 concludes and discusses future work. 2. Overall System Description 2.1. Architecture The INESC-ID IWSLT07 Statistical Machine Translation (SMT) System architecture is shown in Figure 1. It con- sists of a pipeline with the following steps: preprocessing, phrase-based first pass decoding, n-best reranker and post- processing. The first pass module follows the baseline suggested for the ACL second workshop on machine translation1. The used features include both direct and inverse phrase probability, IBM1 model lexica over all possible alignments, and phrase and word penalties. Features are combined using a log linear model optimized to maximize BLEU [2]. In this paper, we focus our description on the remaining modules. 1http://www.statmt.org/wmt07/ Pre-Process Decode Reranker Post-Process Input Output Extra Features Figure 1: System 2.2. Corpora Tables 1 to 3 provide corpora partition and description. One of the characteristics of IWSLT evaluations is the reduced size of the training corpus when compared to other MT eval- uations. Italian English Sentences 19845 Words 14365 184134 Tokens 10126 7011 Avg. Sentence Len. 7.23 9.27 Table 1: Training Corpus From Table 2 to 3 we can see that the average percentage of unknown words is around 10% which represents an addi- tional burden to the already hard task of translation. Also this years task has additional difficulties. First, the source lan- guage (in our case Italian) is lowercase and without punctu- ation; then the corpus is not separated one sentence per line, and there are lines with different number of sentences that must be translated to one another; finally the training cor- pus is not speech transcription, so we have a kind of domain adaptation between the training corpus and the test corpus. 3. Experiments 3.1. Baseline results In order to investigate the difficulty of this year task, a sin- glepass baseline system was trained on the training set and tested on the various development sets. Table 4 shows the results in each set. We can see that this year’s task of sponta- neous speech …

Content truncated to 500 words